基于改进SIFT-ICP算法的Kinect植株点云配准方法

针对传统配准方法准确度低、速度慢的问题,提出了基于改进SIFT-ICP算法的彩色植株点云配准方法。首先采用Kinect获取不同视角下植株彩色图像和深度图像合成原始植株彩色点云,通过预处理提取原始点云植株信息,对植株点云进行尺度不变特征变换(SIFT)的特征点检测,得到点云配准关键点,再对关键点进行自适应法线估计,然后求取关键点的快速点特征直方图(FPFH),通过采样一致性(SAC-IA)初始配准方法改进点云间初始位置关系,最后利用Nanoflann加速最近点迭代(ICP)算法完成精确配准。试验结果表明,改进SIFT-ICP算法可以大幅度提高点云配准的准确性和快速性,其中对应点间平均欧氏距离小于7 mm,配准时间小于30 s。

Method of Plant Point Cloud Registration Based on Kinect of Improved SIFT-ICP

Aiming at solving the low-accuracy and slow-speed problem of traditional registration, an improved SIFT-ICP registration method for color point clouds of plant was put forward. Original color point clouds of plant was merged by color images and depth images obtained by Kinect from different perspectives. Firstly, preprocessing was carried out to extract point clouds of plant from original point clouds, in which lots of point clouds of background and noise were involved. Secondly, by making use of depth features and boundary characteristics of plant point clouds, the key points were detected by means of SIFT (Scale invariant feature transforms) algorithm. Thirdly, normal calculation was executed on the key points computed previously, which was revised by adapting the estimation to accelerate the normal estimation process. The normal estimation was determined by the number of surrounding points. For the sparse part of point cloud, the value of adjacent point was reduced, on the contrary, it was increased in the process of normal estimation. Meanwhile, the FPFH (Fast point feature histograms) descriptor was developed to obtain the characteristic vector which contained 33 dimension element for each key point. Fourthly, SAC-IA (Sample consensus-initial alignment) algorithm, an initial registration algorithm, was applied to register plant color point clouds from different perspectives to provide a better spatial mapping relationships for accurate registration. Finally, on the basis of initial registration, the ICP (Iterative closest point) algorithm, which was accelerated by adapting Nanoflann instead of Flann, was used to refine the initial transform matrix inferred by initial registration. Experiments showed that this registration method can improve not only registration speed but also registration accuracy, the average Euclidean distance between corresponding points was below 7mm and registration time-consuming was less than 30s.